For those who are out of the loop, here is a TLDR of the situation:
Joint statement by the Lean and Mathlib communities, endorsed by the Lean FRO and the Mathlib Initiative.
On November 20, 2026, Levent Alpöge released a proof of the Riemann Hypothesis as a Lean library comprising 17 million lines of code.
We have full confidence that Alpöge’s proof of the Riemann Hypothesis is correct, and we congratulate him on this achievement.
Source: https://leanprover-community.github.io/blog/posts/riemann-hypothesis/
There is several layer to this.
First, just being a stochastic parrot is already really useful in research, it has absorbed all the written human knowledge perhaps several times. Having such a database that is searchable via natural language is already quite valuable.
Second, modern LLM are generally trained two stages, 1) pre-training, which it trys to correctly predict the next word 2) reinforcement training, where it try to solve problem (math, coding, robotics) and its network evolves towards the weight that can solve these problems.
While pre-training gives a scholastic parrot, reinforcement learning is what makes it “useful” (depends on your definition of useful) in math, robotics, and coding. A vast amount of knowledge are stored in natural language, in some sense, the representation of language, i.e. stochastic parrot, unlocked the huge amount of problem to reinforcement learn on. Also the development of formal methods also gives the foundation for these reinforcement: it allows the correctness of result of LLM to be verifiable without human in the loop. (Yet, less than .001% of the AI money have been flowing into ACTUAL formal method research)
AI company are mining (or at least care about) these math problem, because 1) they are easy to mine with verifiable reward 2) they hope intelligence in math can generalize to other domains.
Third, there is an expert human that is skilled in both mathematics and managing LLMs, running with unlimited tokens and bespoke harnesses.
Finally, GRH are really much closer to being solved than we realized, since a human proof is already being reviewed several month ago and confirmed correct by the referee (with no lean certificate). The human proof, of course, is much much cleaner, and have zero AI involvement.
Here are some caveats:
On the math side: unlike what Kevin Buzzard will tell you, correctness is absolutely overrated. It is true all reusable results need to be correct, but mathematics are useful to the world because it creates simple and principled insight to power other fields. That is a “wrong proof of RGH with correct transferable insight” >> “a correct lean slop that takes decades to understand”. That being said, there likely are many insights to be mined from the lean slop, and Levent is commited to do that. But confirming the lean proof is correct is a much much smaller deal than a layman would assume.
On the AI company side: you might wonder why Claude are terrible at archtecturing software yet it can solve GRH. Because the architecture of software don’t really have a good verifiable reward: we archtect software in a way that is easy for human to understand and navigate, there is no way to definitively check that a software is “well-archtected” as opposed to “all test has passed” or “code compiles”.
in the same way, AI produces absolutely terrible math proofs and have a hard time to distill insight, because there is no definitive way to say “what is the spark within this proof”.
Finally, I want to emphasize that “software architecture” and “insight” are likely not unsolvable, just much harder than press a button to see if the test passes. AI company are certainly working on it.
But the real bottleneck, IMO, we do not see that intelligence in math being generalize well to other fields. They still need bespoke training for these models to do accounting, even though most people would assume solving the GRH would imply you can do accounting.
This is likely why they are pushing RSI, because I think both pre-training and reinforcement learning are seeing diminishing return now.
On the money side, AI company are likely burning through more money than all the funding human mathematicians in history combined. Although these model are already here, and inference are cheaper than mathematicians, but people often neglect the mountains of R&D and training cost. So building newer and better LLM just to solve a couple of the most important problems in mathematics might not be a cost-effective deal for anyone.
What is RSI in this context?
“Recursive self improvement”, which like “artificial general intelligence” is not a well-defined term.
Their current pitch is that AI is going to lead to exponential growth of AI research itself, thus they will be able to find the “next big thing” soon.
With grit and determination, like most parrots.
You run your stochastic parrot in a loop generating Lean and have a (proper, deterministic, non-AI) Lean validator run on the output each time. In this case, you also have a human iteratively give the stochastic parrot starting points and general directions to try. And then you burn millions of dollars of compute effectively monkey-typewritering it until the validation passes.
I think people really underestimate the sheer scale of resources allocated to that last part - this is way beyond something a normal mathematician on an academic budget could do and, ironically enough, I don’t think even most mathematicians would argue that this is a sensible use of those resources.
This captures the essence of what the LLM is doing, but is ignoring the fact that the whole thing was being driven by a very good mathematician.
The headlines keep saying things like “Claude proves the Riemann Hypothesis”, but an accurate statement would be “Levent Alpöge proves the Riemann Hypothesis using AI assistance.” The AI didn’t have to search the whole space of possible solutions - Alpöge was there to direct and prune the search space based on his own ideas about what was likely to work. Alpöge couldn’t have done this on his own, but the LLM couldn’t have either.
but is ignoring the fact that the whole thing was being driven by a very good mathematician.
Is it?
In this case, you also have a human iteratively give the stochastic parrot starting points and general directions to try.
Ugh. I swear I read your post three times and someone missed that. My brain is clearly not working this morning. My apologies.
No worries, it’s a fair point about how this keeps getting misattributed to “Claude”
TDLR; they brute forced a wholeass theorem.
I don’t think even most mathematicians would argue that this is a sensible use of those resources.
Is it… because they did the math?
Philosophically, it’s like humanity has figured out a way to automate talking between our past selves.
A human was guiding the parrot, right? I doubt it could solve it itself had someone just written “write a valid proof for the Riemann Hypothesis” and I highly doubt a layman would’ve understood a single thing of what the stochastic machine spit out. People just trust AIs nowadays. “The AI said it, so it must be true”.
My bet is that that, as @HailHydra@infosec.pub alluded to, the solution was already scraped during training. And if not, the power of AI is that they are so fast at testing all the ways not to do something that after enough tries, they should get to a function answer. Again, something that @gnutrino@programming.dev is saying.
It reminds me of a website that can generate anything you want it to from noise or something. It might’ve even be pi.
Edit: found it: The Library Of Babel
Not any human, one that was in the Harvard’s Society of Fellows, has a PhD from Princeton and a masters from Cambridge; got multiple awards and has been exclusively working in solving past mathematical unsolved problems.
He got all the support from a corporation to dedicate to it.
That’s like if we claimed it was Sir Andrew John Wiles’ calculator that resolved Fermat’s Last Theorem: hurrah for the Texas Instruments TI-92!!
Exactly. There no way 99.999% of the planet could’ve successfully guided it.
There’s been multiple examples of this:
https://en.wikipedia.org/wiki/Navier–Stokes_priority_controversy#Background
And more than a few mathematicians have had veiled threats levied at them for speaking up.
AI is desperate to prove shit to juice stock prices. So they find problems humans have been working on, and have solved 99.99999%
In some cases already proven 100%, the AI companies are just comfortable skipping normal vetting/review.
They want the headline, they don’t care if it stands up to scrutiny. A human would have their career ruined rushing shit, AI don’t.
Proving the Riemann Hypothesis is not a problem that humans have proven 99.99999% (except if that one proof turns out to be legit).
Regardless of if it generated a legitimate proof for the Riemann Hypothesis or not, whatever was spat out is less than useless for everyone except maybe Anthropic Shareholders (due to the hype and their name being in the news).
Mathematical proofs are not for their answer, they are for the logical breakthroughs and novel insights that solving a problem like this elicit. The useful extensions to the total collective of human knowledge, that can be reused and applied to other fields to further human development and advancement.
The “proof” that was generated by the LLM is exactly as useful as a napkin with the words “The Riemann Hypothesis is True” written on it. An LLM generated 1.4gb Lean project, and it’s accompanying 250,000 page paper, is so useless to the field of Mathematics that it’s completely insulting to actual insightful work being done. Like the draft paper for an honest human written solution to the Generalised Riemann Hypothesis “The Music Of The Primes” written by Alessandro Maria Masullo who has been working on the paper for over a decade(!) and was less than a year to publication, but had to release it early due to the LLM publicity stunt that is Levent Alpöge/Anthropics “proof”.
Mathematical proofs are in part for their answer. This is true for the Riemann Hypothesis more than any other conjecture, because it is SO big and important that an answer to it immediately solves hundreds of other questions.
In a lot of cases a single theorem is not worth that much, but that’s very much not this case.
And the Riemann Hypothesis is extremely easy to state in Lean, so it’s quite unlikely that a small infidelity has crept into the statement of the theorem, which is the big worry with a huge AI proof - in other words we can be really very confident about the truth of the Riemann Hypothesis now.
Like the draft paper for an honest human written solution to the Generalised Riemann Hypothesis “The Music Of The Primes” written by Alessandro Maria Masullo who has been working on the paper for over a decade(!) and was less than a year to publication, but had to release it early due to the LLM publicity stunt that is Levent Alpöge/Anthropics “proof”.
Lol! I bet the solution to the Generalised Riemann Hypothesis was in the LLM weights due to this.
It’s definitely possible, there are a disturbing number of times where genuine pre-production work has been leaked by an AI company having access to data they shouldn’t.
Though in this instance I don’t believe so as, before the draft was published to the internet, the “The Music Of The Primes” paper had only been shared with one other person, the author’s old Standford Professor/Mentor Brian Conrad. And from the initial surface evaluation of both I’ve heard, it seems the approach the Riemann Hypothesis with completely different methods.
I hate so much of what has come out of AI. I hate that it has done so much damage to the economy. I hate that so many people have lost their jobs, and my own CS degree isn’t worth the paper it’s printed on because entry level no longer exists. I hate that hardware prices are skyrocketing because data centers are selfishly hogging a limited supply of RAM. I hate that those data centers are cooking the planet. I hate that it’s all built on plagiarism and giving nothing back to the people it’s taking everything from. And I hate all the slop being shoved in front of my poor eyeballs.
And that hatred has led me to struggle with processing the fact that despite all the baggage, real things are coming out of this technology. I can’t say that a proof of the most famous open question in mathematics isn’t a big deal. This is the point where I have to admit that the answer to your question is that these models are not just “fancy autocomplete” or “stochastic parrots”, this is real.
I would argue that, yes, it isn’t a big deal.
A mathematical proof does not exist for its answer (or maybe does not solely exist for its answer, depending on the question). It exists for the logical insight and breakthroughs in knowledge that are typically required to answer these famously difficult questions.
An LLM generated 1.4gb Lean project and accompanying 250,000 page paper are completely useless in this regard due to both it’s insane size, and the insanely high ratio of meaningless gaff to actual information that LLMs always generate. So any actual insight that might exist in the proof will almost certainly never be extractable and/or useful to humanity.
If it came up with a cure for cancer that worked but wasn’t able to explain why it works that still wouldn’t be useless. Obviously it would be ideal that it could also teach us but I’m sure there are limits to that as well. There are limitations of human intelligence that do not apply to computers. Not everything even can be understood by a human. For example an AI can tell the person’s gender from a retina picture but even the best eye doctor can’t. They’re both seeing the same picture but the AI is able to detect minute little details that a human may simply be incapable of observing.
But a cure for cancer is useful on its own. A mathematical proof is useful only if it has useful applications (this one doesn’t), or if solving it leads to new math, insights, methods. Which an AI proof doesn’t lead to as it’s just a brute force solution.
If it came up with a cure for cancer that worked but wasn’t able to explain why it works that still wouldn’t be useless
Yeah it would, since you couldn’t reproduce it without knowing how and why it works.
Since LLMs are very prone to hallucinations and only getting WORSE as they’re increasingly being fed data from themself and each other, anything else would be the epitome of irresponsible and unethical behavior.
I feel like you’re comparing apples and oranges here.
There is a vast difference between the solution to an abstract mathematical problem, and a concrete real-world problem.
The cure to cancer would have a direct benefit to humanity, cancer would be cured, less people would die.
The answer to “Is the Reimann Hypothesis true?” does not have a direct benefit, especially if the answer is unverifiable (which an opaque insanely large blob of a program and paper most certainly are). The benifit of a solution to that question lie in both the cascading effect of it verifying other mathematical proofs that assume the Reimann Hypothesis is true, and the new insights that are developed in the process of solving it.
if the answer is unverifiable (which an opaque insanely large blob of a program and paper most certainly are).
Just to clarify, the LEAN system used for writing up the proof is exactly meant for verifiability, and is pretty solid as I understand it. It’s built to programmatically verify the logic consistency and truth of each step of the proof, which is sufficient to verify the total proof.
So any actual insight that might exist in the proof will almost certainly never be extractable and/or useful to humanity.
Why? It’ll take a long time but people will almost certainly be going through this proof. It might be bullshit or useless, but it might not be if the actual information can be usefully extracted.
I hear you.
I’m anti AI, but anti AI culture has kinda become a Reddit style bubble, I think.
The stochastic parrot thing made sense as a critique against notions that these were fully conscious and embodied thoughtful creatures of intelligence.
But I don’t think it ever made sense as a complete explanation for what they are. Their whole point is to scale the training data and model size up to the point that their “stochastic” behaviour becomes something beyond just predicting. Because if something can predict well enough and deeply enough and from a long enough starting point (IE prompt), well what is intelligence again? Not the same thing? Sure. But probably something new.
To go on crying that AIs are just dumb useless machines is a good way to lose the attention of everyone that does find them useful in some way.
So why be anti AI? All the other things other than whether they are intelligent right now. I’m constantly surprised at how much people struggle to talk about and how much being against a form of technological progress is taboo (almost like a cult/religion).
I use AI quite a bit and there are subtle moments every now and then that it reminds me of how there’s more going on under the hood than just parroting what it has read before. For example I was recently asking about the compatibility of a certain indoor primer to be applied under outdoor cement product and in normal use case these do not work together and I know from experience that it advises against that because I’ve asked before. However this time it said it was fine because the reason they do not work together outside doesn’t apply in this particular use case. So not only does it know what the product sheet tells is the right way to use the product, it’s also able to look at the broader picture and understand when those limitations don’t apply.
Did you double check the product sheets to make sure it didn’t just fabricate that advice?
This is off-label use for said products. The products I used were Weber MD16 Dispersion and Weber Outdoor Repair cement. In the past when I was using the cement for repairing outdoor stairs I asked about priming with the MD16 and it told me not to because it also creates a waterproof layer which then prevents any moisture in the concrete from evaporating and causes delamination. That is correct information according to the product sheet. These are not to be used together, one is an indoor product and the other is an outdoor one.
However, in the latter case I’m applying it over a model scenery made out of cardboard. There cardboard offers a poor adhesion and it would immediately suck any moisture out of the cement making it dry too quickly. Here it told me to absolutely use the MD16 because it fixes both of those issues. However, this is not something the product sheet talks about. If you were to just blindly stare at that with no consideration for the broader context the answer would be to not combine the two. However I did and it worked out perfectly.
Quick answer: “Fancy autocomplete” and “stochastic parrot” are dismissive oversimplifications of what these programs are doing. There is more going on. Whether they have “true intelligence” (whatever that is) is more of a philosophical question than a mathematical one by now.
This article is almost a year old and shows that “fancy autocomplete” was either already history or never true: https://thegradient.pub/othello/
The fancy auto complete thing was always bullshit. Because for most people their first prolonged contact with LLMs was copilot in vs code so they missed the forest for the trees and mistook it for some intellisense thing.
This and the “it just calculates the probability of the next token” simplification completely ignore what makes transformers with attention heads groundbreaking.
It’s not a stochastic parrot or autocomplete. Well, not just those things. Yes, at their core LLMs are still just predicting tokens, but there’s been a lot of work on harness software and model improvements (though it blows my mind that they can apparently generate mathematical proofs while still being unable to consistently count the instances of “r” in “strawberry.”)
That said, I have my doubts. In a Lean proof that long, there’s ample opportunity to find and exploit unknown flaws in the kernel, and no human has had the time to check it (let alone the absurd 250,000 pages of the argument in prose).
I’m going to withhold judgment until an actual team of mathematicians can verify it.
Yeah it’s crazy but I think at this level of math a lot of the proving happens in concept space and not necessarily in number space.
It cannot count letters, but it can write a script that does it. Chatbots are toys, llms are stochastic parrots, but a well built agent using llms to write code is very capable. And I am not a mathematician but my hunch is that the AI companies basically treat these math problems as coding assignments. They are not looking for new discoveries, they want to code proofs for theorems that mathematicians formulated.
It cannot count letters, but it can write a script that does it.
Sure. Little one-off scripts like that, they’re very good at.
Chatbots are toys, llms are stochastic parrots, but a well built agent using llms to write code is very capable.
Eh. I think this is less a function of them being capable and more a function of how much money you’re willing to spend to brute-force it. A limited token budget severely restricts the value of LLM coding, because no matter how good the harness and how carefully considered the prompt, it will inevitably make mistakes, and if you run out of tokens those mistakes are extremely costly.
If money is no object, you can force your harness to keep working on a problem until it eventually stumbles on the solution (or an exploit that allows it to appear like it’s found the solution). But nobody has an endless budget for that.
And I am not a mathematician but my hunch is that the AI companies basically treat these math problems as coding assignments. They are not looking for new discoveries, they want to code proofs for theorems that mathematicians formulated.
Agreed, that’s exactly what they’re doing. And that’s why I’m withholding judgment until there’s a human team of mathematicians who can verify the results. Lean is great, but it’s not perfect, and there’s ample opportunity in a multi-million line Lean proof for errors that the Lean kernel doesn’t catch.
You know there’s more to ai than large language models right?
Yeah… the implication that all AI is LLMs is ridiculous. There is super legitimate real AI making a huge difference in tons of fields right now.
I mean, sure. Neural networks existed for longer than llms, but in this case they literally used an llm so quit being outraged out of nothing
I mean objections to their uses aside, LLMs are super legitimate real AI. Seriously these things are pretty insane, and there has been some good work done with them in many fields.
I don’t honestly know, but LLMs are very good at breaking the rules to technically achieve their goals. I don’t know why someone would assume 17 million lines of code don’t have some kind of novel exploit.
It’s been tested with multiple verifiers and although no humans currently understand all the proof’s low level details, the high level approach has been understood to some extent, and makes sense. The proof is almost certainly valid. Whether it was worth doing, represents an advance in math, and all that sort of thing are separate questions of course.
Just as an extension of this, Lean has had proof invalidating bugs exploited by LLMs.
fancy autocomplete / stochastic parrot is just the user interface. what they can do now (so far kinda poorly but improving) is much more than when it all started.
I’m no expert at all …
But I think the secret sauce that hasn’t quite made it into mainstream understanding is this:
It’s not one parrot any more. It’s a whole bunch working together with different assigned roles and prompts feeding one output into another as input and so on. Not unlike how society and the brain work.
Combine that with external systems that the models can use as a metric or a goal (such as “make this program return success”), and you have something not unlike evolution where the models and their inputs and outputs to each other are like mutations over generations, the external system like natural selection, and the models themselves like all the physics and biochemistry that is harnessed by life.










